This repo maintains a canonical portable autoresearch loop in loop.md: a reference abstraction that others can adapt to specific domains and implementations.
Its main reference case is karpathy/autoresearch, but the goal here is not to mirror that repo. The goal is to extract the underlying loop architecture, keep it portable, and refine it when new evidence justifies a change.
This repo is deliberately small and only really has three tracked files that matter:
program.md- the agent control surface and daily run instructionsloop.md- the canonical abstraction and porting guideREADME.md- the repo contract
Local untracked artifacts:
ledger.md- append-only local run log.scratch/github-inspection/- temporary inspection area
program.md is the automation entry point. Each scheduled run completes one observe cycle and stops. Only update now runs may enter a bounded promotion loop over loop.md; recurrence comes from the scheduler.
ledger.md is local run state and an append-only ledger for future runs, not part of the promoted canonical doc surface. ledger.md and .scratch/ stay out of git.
This repo tracks portable architecture changes. The examples below are illustrative, not a binding checklist:
- control surfaces
- mutable versus fixed boundaries
- verification and scorekeeping
- crash handling
- unattended-run ergonomics
- artifact logging and git hygiene
- coordination layers above the core loop
- platform, backend, or metric shifts that create a genuinely different loop
If a signal is interesting only because of the original training domain, it is out of scope here.
See program.md for the run procedure and loop.md for the abstraction being maintained.
## License
MIT
Cavit Erginsoy, 2026